2003-09-02 Dave Patterson (cs.berkeley/~patterson) www-inst.eecs.berkeley/~cs152/

1. CS152 – Computer Architecture andEngineeringLecture 3 – Field Programmable Gate Arrays 2003-09-02 Dave Patterson (www.cs.berkeley.edu/~patterson) www-inst.eecs.berkeley.edu/~cs152/

2. Review: Verilog • Verilog allows both structural and behavioral descriptions, helpful in testing • Some special features only in Hardware Description Languages • # time delay, nonblocking assignments, initial vs. always, forever loops • Syntax a mixture of C (operators, for, while, if, print) and Ada (begin… end, case…endcase, module …endmodule) • Verilog can describe everything from single gate to full computer system; you get to design a simple processor

3. Multiple Review • Multiply: successive refinement to see final design • 1st iteration: 64-bit Adder, 64-bit Multiplicand shift register, 32-bit Multiplier shit register, 64-bit Product register • 3rd iteration: 32-bit Adder, 64-bit Product/Mutiplier shift register, 32-bit Multiplicand Register • There are algorithms that calculate many bits of multiply per cycle (see exercises 4.36 to 4.39 in COD)

4. Outline • FPGAs Overview • Why use FPGAs?(a short history lesson). • FPGA variations • Internal logic blocks. • Designing with FPGAs. • Specifics of Xilinx Virtex-E series.

5. FPGA Overview • Basic idea: 2D array of combination logic blocks (CL) and flip-flops (FF) with a means for the user to configure both: 1. the interconnection between the logic blocks, 2. the function of each block. Simplified version of FPGA internal architecture

6. Why FPGAs? (1 / 5) • By the early 1980’s most of logic circuits in typical systems were absorbed by a handful of standard large scale integrated circuits (LSI ICs). • Microprocessors, bus/IO controllers, system timers, ... • Every system still needed random small “glue logic” ICs to help connect the large ICs: • generating global control signals (for resets etc.) • data formatting (serial to parallel, multiplexing, etc.) • Systems had a few LSI components and lots of small low density SSI (small scale IC) and MSI (medium scale IC) components. Printed Circuit (PC) board with many small SSI and MSI ICs and a few LSI ICs

7. Why FPGAs? (2 / 5) • Custom ICs sometimes designed to replace glue logic: • reduced complexity/manufacturing cost, improved performance • But custom ICs expensive to develop, and delay introduction of product (“time to market”) because of increased design time • Note: need to worry about two kinds of costs: 1. cost of development, “Non-Recurring Engineering (NRE)”, fixed 2. cost of manufacture per unit, variable Usually tradeoff between NRE cost and manufacturing costs NRE NRE

8. Why FPGAs? (3 / 5) • Therefore custom IC approach was only viable for products with very high volume (where NRE could be amortized), and not sensitive in time to market (TTM) • FPGAs introduced as alternative to custom ICs for implementing glue logic: • improved PC board density vs. discrete SSI/MSI components (within around 10x of custom ICs) • computer aided design (CAD) tools meant circuits could be implemented quickly (no physical layout process, no mask making, no IC manufacturing), relative to Application Specific ICs (ASICs) (3-6 months for these steps for custom IC) • lowers NREs (Non Recurring Engineering) • shortens TTM (Time To Market) • Because of Moore’s law the density (gates/area) of FPGAs continued to grow through the 80’s and 90’s to the point where major data processing functions can be implemented on a single FPGA.

9. Why FPGAs? (4 / 5) • FPGAs continue to compete with custom ICs for special processing functions (and glue logic) but now try to compete with microprocessors in dedicated and embedded applications • Performance advantage over microprocessors because circuits can be customized for the task at hand. Microprocessors must provide special functions in software (many cycles) • MICRO: Highest NRE, SW: fastest TTM • ASIC: Highest performance, worst TTM • FPGA: Highest cost per chip (unit cost)

10. Why FPGAs? (5 / 5) • As Moore’s Law continues, FPGAs work for more applications as both can do more logic in 1 chip and faster • Can easily be “patched” vs. ASICs • Perfect for courses: • Can change design repeatedly • Low TTM yet reasonable speed • With Moore’s Law, now can do full CS 152 project easily inside 1 FPGA

11. Administrivia • Prerequisite Quiz Results • Lab 1 due tomorrow • How many bought $37 PRS Transmitor ?from behind ASUC textbook desk (Chem 1A, CS 61ABC, 160) • Can sell back to bookstore

12. Where are FPGAs in the IC Zoo? Logic Standard Logic ASIC CPLDs Source: Dataquest Programmable Logic Devices (PLDs) Gate Arrays Cell-Based ICs Full Custom ICs SPLDs (PALs) FPGAs Acronyms SPLD = Simple Prog. Logic Device PAL = Prog. Array of Logic CPLD = Complex PLD FPGA = Field Prog. Gate Array (Standard logic is SSI or MSI buffers, gates) • Common Resources • Configurable Logic Blocks (CLB) • Memory Look-Up Table • AND-OR planes • Simple gates • Input / Output Blocks (IOB) • Bidirectional, latches, inverters, pullup/pulldowns • Interconnect or Routing • Local, internal feedback, and global

13. Families of FPGA’s differ in: physical means of implementing user programmability, arrangement of interconnection wires, and basic functionality of logic blocks Most significant difference is in the method for providing flexible blocks and connections: Anti-fuse based (ex: Actel) Non-volatile, relatively small - fixed (non-reprogrammable) (Almost used in 150 Lab: only 1-shot at getting it right!) FPGA Variations

14. Latches are used to: 1. make or break cross-point connections in interconnect 2. define function of logic blocks 3. set user options: within the logic blocks in the input/output blocks global reset/clock “Configuration bit stream” loaded under user control: All latches are strung together in a shift chain “Programming” => creating bit stream User Programmability • Latch-based (Xilinx, Altera, …) • reconfigurable - volatile • relatively large die size • Note: Today 90% die is interconnect, 10% is gates

15. Idealized FPGA Logic Block • 4-input Look Up Table (4-LUT) • implements combinational logic functions • Register • optionally stores output of LUT • Latch determines whether read reg or LUT

16. 4-LUT Implementation • n-bit LUT is actually implemented as a 2n x 1 memory: • inputs choose one of 2n memory locations. • memory locations (latches) are normally loaded with values from user’s configuration bit stream. • Inputs to mux control are the CLB (Configurable Logic Block) inputs. • Result is a general purpose “logic gate”. • n-LUT can implement any function of n inputs!

17. An n-lut as a direct implementation of a function truth-table Each latch location holds value of function corresponding to one input combination LUT as general logic gate Example: 4-lut Example: 2-lut Implements any function of 2 inputs. How many functions of n inputs?

18. More functionality for “free”? • Given basic idea • LUT built from RAM • Latches connected as shift register • What other functions could be provided at very little extra cost? • Using CLB latches as little RAM vs. logic • Using CLB latches as shift register vs. logic

19. 1. “Distributed RAM” RAM16X1S D WE WCLK = O A0 A1 A2 A3 LUT LUT LUT RAM32X1S D WE WCLK A0 O A1 A2 A3 A4 or RAM16X2S D0 D1 WE = WCLK O0 A0 O1 RAM16X1D A1 A2 D A3 WE or WCLK A0 SPO A1 A2 A3 DPRA0 DPO DPRA1 DPRA2 DPRA3 • CLB LUT configurable as Distributed RAM • A LUT equals 16x1 RAM • Implements Single and Dual-Ports • Cascade LUTs to increase RAM size • Synchronous write • Synchronous/Asynchronous read • Accompanying flip-flops used for synchronous read

20. 2. Shift Register LUT D D D D Q Q Q Q IN CE CE CE CE CE CLK LUT = OUT DEPTH[3:0] • Each LUT can be configured as shift register • Serial in, serial out • Saves resources: can use less than 16 FFs • Faster: no routing • Note: CAD tools determine with CLB used as LUT, RAM, or shift register, rather than up to designer

21. How Program: FPGA Generic Design Flow • Design Entry: • Create your design files using: • schematic editor or • hardware description language (Verilog, VHDL) • Design “implementation” on FPGA: • Partition, place, and route (“PPR”) to create bit-stream file • Divide into CLB-sized pieces, place into blocks, route to blocks • Design verification: • Use Simulator to check function, • Other software determines max clock frequency. • Load onto FPGA device (cable connects PC to board) • check operation at full speed in real environment.

22. Example Schematic Circuit: collection of gates and flip-flops Idealized FPGA structure: Example Partition, Placement, and Route Circuit combinational logic must be “covered” by 4-input 1-output “gates”. Flip-flops from circuit must map to FPGA flip-flops. (Best to preserve “closeness” to CL to minimize wiring.) Placement in general attempts to minimize wiring.

23. Xilinx Vittex-E Routing Hierarchy INTERNAL BUSSES Internal 3-state Bus Long lines and Global lines Buffered Hex lines (1/6 blocks) Note: CAD tools do PPR, not designers Single-length lines DIRECT CONNECTION Direct connections • 24 single-length lines • Route GRM signals to adjacent GRMs in 4 directions • 96 buffered hex lines • Route GRM (general routing matrix) signals to another GRMs six blocks away in each of the 4 directions • 12 buffered Long lines • Routing across top and bottom, left and right

24. 2 “logic slices” / CLB, two 4-LUTs / slice=> Four 4-LUTs / CLB Virtex-E Configurable Logic Block (CLB)

25. Peer Instruction • How would you place ASIC, FPGA, and Microprocessors+software from best to worst? • Performance? • Non Recurring Engineering? • Unit cost? • Time To Market? • ASIC, FPGA, MICRO • ASIC, MICRO, FPGA • FPGA, ASIC, MICRO • FPGA, MICRO, ASIC • MICRO, ASIC, FPGA • MICRO, FPGA, ASIC

26. Virtex-E CLB Slice Structure • Each slice contains two sets of the following: • Four-input LUT • Any 4-input logic function • Or 16-bit x 1 sync RAM • Or 16-bit shift register • Carry & Control • Fast arithmetic logic • Multiplexer logic • Multiplier logic • Storage element • Latch or flip-flop • Set and reset • True or inverted inputs • Sync. or async. control

27. Details of Virtex-E Slice Very fast ripple carry:(24-bit @ 100 MHz) Multiplexors to helpcombine CLBs into larger multiplexor

28. CLB Slice MUXF6 MUXF5 Slice MUXF5 LUT LUT LUT LUT Virtex-E Dedicated Expansion Multiplexers • Since 4-LUT has 4 inputs, max is 2:1 Mux (2 inputs, 1 control line) • MUXF5 combines 2 LUTs to create • 4x1 multiplexer • Or any 5-input function (5-LUT) • Or selected functions up to 9 inputs • MUXF6 combines 2 slices to form • 8x1 multiplexer • Or any 6-input function (6-LUT) • Or selected functions up to 19 inputs • Dedicated muxes are faster and more space efficient

29. Xilinx Virtex-E Chip Floorplan • Input / Output Blocks (IOBs) • Configurable Logic Blocks (CLBs) • Block RAMs (BRAMs) (discussed soon) • Delay Locked Loop (DLL) (discussed soon) • “VersaRing” =

30. Block RAM (Extra RAM not using LUTs) Port A Spartan-IIE True Dual-Port Block RAM Port B Block RAM • Most efficient memory implementation • Dedicated blocks of memory • Ideal for most memory requirements • Virtex-E XCV2000 has 160? blocks • 4096 bits per blocks • Use multiple blocks for larger memories • Builds both single and true dual-port RAMs • CORE Generator provides custom-sized block RAMs • Quickly generates optimized RAM implementation

31. Virtex-E Block RAM • Flexible 4096-bit block… Variable aspect ratio • 4096 x 1 • 2048 x 2 • 1024 x 4 • 512 x 8 • 256 x 16 • Increase memory depth or width by cascading blocks

32. Clock De-skew Virtex-E Delay Lock Loop (DLL) Capabilities • Easy clock duplication • System clock distribution • Cleans and reconditions incoming clock • Quick and easy frequency adjustment • Single crystal easily generates multiple clocks • Excellent for advance memory types • De-skew incoming clock • Generate fast setup and hold time or fast clock-to-outs

33. 66MHz - 2x Clock Multiplication DLL 132 MHz (Multiply by 2) 66 MHz DLL: Multiplication of Clock Speed • Have faster internal clock relative to external clock source • Use 1 DLL for 2x multiplication • Combine 2 DLLs for 4x multiplication • Reduce board EMI • Route low-frequency clock externally and multiply clock on-chip

34. 180° Phase Shift 30 MHz (180° Shift) DLL 30 MHz 30 MHz Used for FB DLL 15 MHz (Divide by 2) 30 MHz (180° Shift) 60 MHz (Multiply by 2) Clock x2 and Clock 2 DLL: Division of Clock Speed • Selectable division values • 1.5, 2, 2.5, 3, 4, 5, 8, or 16 • Cascade DLLs to combine functions • Combine DLLs to multiply and divide to get desired speed • 50/50 duty cycle correction available

35. Clock Management Summary • All digital DLL Implementation • Input noise rejection • 50/50 duty cycle correction • Clock mirror provides system clock distribution • Multiply input clock by 2x or 4x • Divide clock by 1.5, 2, 2.5, 3, 4, 5, 8, or 16 • De-skew clock for fast setup, hold, or clock-to-out times

36. Virtex-E Family of Parts

37. Summary: Xilinx FPGAs • How they differ from idealized array: • In addition to their use as general logic “gates”, LUTs can alternatively be used as general purpose RAM or shift register • Each 4-LUT can become a 16x1-bit RAM array • Special circuitry to speed up “ripple carry” in adders and counters • Therefore adders assembled by the CAD tools operate much faster than adders built from gates and LUTs alone. • Many more wires, including tri-state capabilities.

38. Backup Slides FYI

39. 3 Operand Adder Function COUT COUT Carry & Control Logic Carry & Control Logic Look-Up Table Look-Up Table PARTIAL1 O B1 B0 SUM1 Look-Up Table Carry & Control Logic Look-Up Table Carry & Control Logic PARTIAL0 A1 A0 SUM0 CIN CIN SLICE0 SLICE1 CLB C1 C0 • A, B, C are two-bits wide • SUM = A + B + C or PARTIAL + C, where PARTIAL = A + B • Implementation • First 2-operand sum ‘A+B’ is performed in Slice 0 • Second 2-operand sum ‘PARTIAL + C’ is performed in Slice 1 • Fast local feedback connection within the CLB • Very small delay for on PARTIAL

40. Dedicated CLB Multiplier Logic Dedicated AND gate • Dedicated AND gate • Highly efficient ‘Shift & Add’ implementation • For a 16x16 Multiplier • 30% reduction in area and one less logic level

41. Xilinx FPGAs (interconnect detail)

42. Virtex-E Input/Output block (IOB) detail